Add Docker deployment support

- Add -y flag to ffmpeg to overwrite mkstemp-created empty temp files
  (ffmpeg exits 0 without writing when file exists and stdin is /dev/null)
- Fix language fallback: use `or` instead of dict.get default so empty
  string from client falls back to config value
- Fix language value in config: ru -> russian

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.dockerignore

@@ -0,0 +1,17 @@
+.git
+.gitignore
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+.Python
+.pytest_cache/
+.mypy_cache/
+.venv/
+venv/
+env/
+logs/
+history/
+models/
+tmp/
+client.png

Dockerfile

@@ -0,0 +1,27 @@
+FROM python:3.12-slim
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    DEBIAN_FRONTEND=noninteractive \
+    PIP_NO_CACHE_DIR=1
+
+WORKDIR /app
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        ffmpeg \
+        sox \
+        libmagic1 \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt ./
+RUN pip install --upgrade pip \
+    && pip install -r requirements.txt
+
+COPY . .
+
+RUN mkdir -p /app/logs /app/history
+
+EXPOSE 5042
+
+CMD ["python", "server.py", "--config", "config.docker.json"]

README.md

@@ -63,6 +63,56 @@ pip install -r requirements.txt
 python server.py
 ```
 
+## Docker
+
+The repository now includes a container setup for running the API in Docker.
+
+### What gets mounted
+
+- `./config.docker.json` -> `/app/config.docker.json`
+- `./logs` -> `/app/logs`
+- `./history` -> `/app/history`
+- `./models` -> `/models`
+
+Place your Whisper model inside `./models/whisper` or update `model_path` in `config.docker.json`.
+
+### Build and run with Docker Compose
+
+```bash
+docker compose up --build
+```
+
+The API will be available at:
+
+```text
+http://localhost:5042
+```
+
+### Build and run with plain Docker
+
+```bash
+docker build -t whisper-api-server .
+docker run --rm -p 5042:5042 \
+  -v "$(pwd)/config.docker.json:/app/config.docker.json:ro" \
+  -v "$(pwd)/logs:/app/logs" \
+  -v "$(pwd)/history:/app/history" \
+  -v "$(pwd)/models:/models" \
+  whisper-api-server
+```
+
+### GPU note
+
+The container image installs the same Python dependencies as the local setup, including CUDA-oriented PyTorch wheels on Linux x86_64. To actually use NVIDIA GPU acceleration at runtime, start the container with GPU access enabled in your Docker environment, for example:
+
+```bash
+docker run --rm --gpus all -p 5042:5042 \
+  -v "$(pwd)/config.docker.json:/app/config.docker.json:ro" \
+  -v "$(pwd)/logs:/app/logs" \
+  -v "$(pwd)/history:/app/history" \
+  -v "$(pwd)/models:/models" \
+  whisper-api-server
+```
+
 ## Configuration
 
 The service is configured through the `config.json` file:
@@ -73,14 +123,26 @@ The service is configured through the `config.json` file:
     "model_path": "/path/to/whisper/model",
     "language": "russian",
     "enable_history": true,
+    "max_history_days": 30,
     "chunk_length_s": 28,
-    "batch_size": 8,
+    "batch_size": 6,
     "max_new_tokens": 384,
     "temperature": 0.01,
     "return_timestamps": false,
-    "audio_rate": 8000,
+    "audio_rate": 16000,
     "norm_level": "-0.55",
-    "compand_params": "0.3,1 -90,-90,-70,-50,-40,-15,0,0 -7 0 0.15"
+    "compand_params": "0.3,1 -90,-90,-70,-50,-40,-15,0,0 -7 0 0.15",
+    "device_id": 0,
+    "file_validation": {
+        "max_file_size_mb": 500,
+        "allowed_extensions": [".wav", ".mp3", ".ogg", ".flac", ".m4a", ".oga", ".aac", ".webm"],
+        "allowed_mime_types": ["audio/wav", "audio/mpeg", "audio/ogg", "audio/flac", "audio/mp4", "audio/x-m4a", "audio/aac", "audio/webm"]
+    },
+    "log_level": "INFO",
+    "log_file": "logs/whisper_api.log",
+    "request_logging": {
+        "exclude_endpoints": ["/health", "/static"]
+    }
 }
 ```
 
@@ -92,6 +154,7 @@ The service is configured through the `config.json` file:
 | `model_path` | Path to the Whisper model directory |
 | `language` | Language for transcription (e.g., "russian", "english") |
 | `enable_history` | Whether to save transcription history (true/false) |
+| `max_history_days` | Number of days to keep transcription history before rotation |
 | `chunk_length_s` | Length of audio chunks for processing (in seconds) |
 | `batch_size` | Batch size for processing |
 | `max_new_tokens` | Maximum new tokens for the model output |
@@ -100,6 +163,13 @@ The service is configured through the `config.json` file:
 | `audio_rate` | Audio sampling rate in Hz |
 | `norm_level` | Normalization level for audio preprocessing |
 | `compand_params` | Parameters for audio compression/expansion |
+| `device_id` | CUDA device index to use for inference |
+| `file_validation.max_file_size_mb` | Maximum allowed file size in megabytes |
+| `file_validation.allowed_extensions` | List of accepted audio file extensions |
+| `file_validation.allowed_mime_types` | List of accepted MIME types |
+| `log_level` | Logging level (DEBUG, INFO, WARNING, ERROR) |
+| `log_file` | Path to the log file |
+| `request_logging.exclude_endpoints` | Endpoints excluded from request logging |
 
 ## Web interface
 
@@ -159,14 +229,33 @@ curl -X POST http://localhost:5042/v1/audio/transcriptions/base64 \
   -d '{"file":"base64_encoded_audio_data"}'
 ```
 
-### Transcribe a local file on the server
+### Transcribe asynchronously
+
+Submit a file for background transcription and receive a task ID:
 
 ```bash
-curl -X POST http://localhost:5042/local/transcriptions \
+curl -X POST http://localhost:5042/v1/audio/transcriptions/async \
-  -H "Content-Type: application/json" \
+  -F file=@audio.mp3
-  -d '{"file_path":"/path/to/audio.mp3"}'
 ```
 
+Response:
+```json
+{"task_id": "abc123..."}
+```
+
+### Get async task status
+
+```bash
+curl http://localhost:5042/v1/tasks/<task_id>
+```
+
+Response when completed:
+```json
+{"task_id": "abc123...", "status": "completed", "result": {...}}
+```
+
+Possible statuses: `pending`, `completed`, `failed`.
+
 ### Request with additional parameters
 
 ```bash
@@ -215,45 +304,22 @@ curl -X POST http://localhost:5042/v1/audio/transcriptions \
 }
 ```
 
-## Project structure
-
-The project consists of the following components:
-
-- `server.py`: Entry point that initializes and starts the service
-- `server.sh`: Bash script for launching the server with optional conda environment update
-- `config.json`: Service configuration file
-- `app/`: Main application module
-  - `__init__.py`: Contains the `WhisperServiceAPI` class for service initialization
-  - `routes.py`: API route definitions
-  - `history.py`: Saving transcription history
-  - `core/`: Core logic
-    - `transcriber.py`: `WhisperTranscriber` class for speech recognition
-    - `transcription_service.py`: Manages the transcription workflow
-  - `audio/`: Audio processing
-    - `processor.py`: `AudioProcessor` class for audio preprocessing
-    - `sources.py`: Audio source handlers (upload, URL, base64)
-    - `utils.py`: Audio utilities (loading, duration)
-  - `infrastructure/`: Supporting modules
-    - `log.py`: Logging configuration
-    - `validation.py`: File validation
-    - `storage.py`: Temp file management
-    - `async_tasks.py`: Async task manager
-  - `static/`: Web interface files
-
 ## Advanced usage
 
 ### Using with different models
 
 You can use any Whisper model by changing the `model_path` in the configuration:
 
-1. Download a model from Hugging Face (e.g., `openai/whisper-large-v3`)
+1. Download a model from Hugging Face
 2. Update the `model_path` in `config.json`
 3. Restart the service
 
+The recommended model for Russian speech recognition is [whisper-large-v3-russian-ties-podlodka-v1.2](https://huggingface.co/Apel-sin/whisper-large-v3-russian-ties-podlodka-v1.2).
+
 ### Hardware acceleration
 
 The service automatically selects the best available compute device:
-- CUDA GPU (index 1 if available, otherwise index 0)
+- CUDA GPU (device index configured via `device_id` in `config.json`)
 - Apple Silicon MPS (for Mac with M1/M2/M3 chips)
 - CPU (fallback)
 

app/audio/processor.py

@@ -59,6 +59,7 @@ class AudioProcessor:
             "ffmpeg",
             "-hide_banner",
             "-loglevel", "warning",
+            "-y",
             "-i", input_path,
             "-ar", f"{audio_rate}",
             "-ac", "1",  # Монофонический звук

app/core/transcriber.py

@@ -143,13 +143,23 @@ class WhisperTranscriber:
             use_safetensors=True,
         )
 
+        use_flash_attn = False
+        if self.device.type == "cuda":
+            # Flash Attention 2 требует архитектуру Ampere или новее (compute capability >= 8.0)
+            capability = torch.cuda.get_device_capability(self.device.index)
+            if capability[0] >= 8:
+                use_flash_attn = True
+                logger.info("GPU поддерживает Flash Attention 2 (compute capability: %d.%d)", *capability)
+            else:
+                logger.info("GPU не поддерживает Flash Attention 2 (compute capability: %d.%d), используется стандартный режим", *capability)
+
         try:
-            if self.device.type == "cuda":
+            if use_flash_attn:
                 model_kwargs["attn_implementation"] = "flash_attention_2"
             self.model = WhisperForConditionalGeneration.from_pretrained(
                 self.model_path, **model_kwargs
             ).to(self.device)
-            if self.device.type == "cuda":
+            if use_flash_attn:
                 logger.info("Используется Flash Attention 2")
         except Exception as e:
             logger.warning("Не удалось загрузить модель с Flash Attention: %s", e)

app/core/transcription_service.py

@@ -36,7 +36,7 @@ class TranscriptionService:
             Кортеж (JSON-ответ, HTTP-код).
         """
         params = params or {}
-        language = params.get('language', self.config.get('language', 'en'))
+        language = params.get('language') or self.config.get('language', 'en')
         temperature = max(0.0, min(1.0, float(params.get('temperature', 0.0))))
         prompt = params.get('prompt', '')
 

config.docker.json

@@ -0,0 +1,26 @@
+{
+    "service_port": 5042,
+    "model_path": "/models/whisper",
+    "language": "russian",
+    "enable_history": true,
+    "max_history_days": 30,
+    "chunk_length_s": 28,
+    "batch_size": 6,
+    "max_new_tokens": 384,
+    "temperature": 0.01,
+    "return_timestamps": false,
+    "audio_rate": 16000,
+    "norm_level": "-0.55",
+    "compand_params": "0.3,1 -90,-90,-70,-50,-40,-15,0,0 -7 0 0.15",
+    "device_id": 0,
+    "file_validation": {
+        "max_file_size_mb": 500,
+        "allowed_extensions": [".wav", ".mp3", ".ogg", ".flac", ".m4a", ".oga", ".aac", ".webm"],
+        "allowed_mime_types": ["audio/wav", "audio/mpeg", "audio/ogg", "audio/flac", "audio/mp4", "audio/x-m4a", "audio/aac", "audio/webm"]
+    },
+    "log_level": "INFO",
+    "log_file": "logs/whisper_api.log",
+    "request_logging": {
+        "exclude_endpoints": ["/health", "/static"]
+    }
+}

config.json

@@ -1,7 +1,7 @@
 {
     "service_port": 5042,
     "model_path": "/home/text-generation/models/whisper/podlodka-turbo",
-    "language": "ru",
+    "language": "russian",
     "enable_history": true,
     "max_history_days": 30,
     "chunk_length_s": 28,

docker-compose.yml

@@ -0,0 +1,14 @@
+services:
+  whisper-api:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    container_name: whisper-api-server
+    ports:
+      - "5042:5042"
+    volumes:
+      - ./config.docker.json:/app/config.docker.json:ro
+      - ./logs:/app/logs
+      - ./history:/app/history
+      - ./models:/models
+    restart: unless-stopped